Wikantik already has a retrieval experiment harness; what's pending is the scheduler that turns it into a nightly CI gate. Until Phase 5 ships, the harness is invokable manually for spot-checks.
Before merging a retrieval-touching change, or when investigating an anecdotal regression.
The harness lives in wikantik-knowledge/src/main/java/com/wikantik/knowledge/eval/
(or its current equivalent). It evaluates a query set against a chosen
retrieval mode, computes nDCG@k, Recall@k, and MRR, and returns
per-query + aggregate scores.
Phase 5 of AgentGradeContentDesign adds the missing pieces: a
scheduled runner, a database-backed query-set store
(retrieval_query_sets / retrieval_queries / retrieval_runs),
Prometheus gauges, and a RetrievalQualitySmokeTest that runs in CI
on every merge.
The frontmatter steps are the canonical procedure for the manual
path. Once Phase 5 lands, the same logic runs without operator
intervention nightly.
The frontmatter pitfalls capture the recurring methodology mistakes.
"One run is authoritative" is the most common — retrieval scores have
real variance across runs and any single number is suspect.